Air Canada was ordered to pay C$812.02 after a chatbot contradicted its own fare policy.

In November 2022, Jake Moffatt was arranging travel after a family bereavement. The airline’s chatbot said a reduced bereavement fare could be requested within 90 days of the ticket being issued, including after travel. The policy page linked from that answer said requests after completed travel were excluded.

Moffatt relied on the chatbot when booking and applied within its stated window. On 14 February 2024, the tribunal ordered Air Canada to pay C$812.02: C$650.88 in damages, C$36.14 in interest and C$125 in fees.

The tribunal did not identify the technology behind the chatbot.

Somewhere in your own material, could a customer find the same kind of contradiction? The refund policy says one thing. The campaign says another. Both carry your name.

Would your customers approve what your team approved?

An internal review can feel reassuring. Everyone knows the product, understands the shorthand and remembers the condition that was dropped from the latest draft. A first-time visitor has none of that background.

Research from IAB and Sonata Insights, published in January 2026, gives marketers another reason to look outside the room. Among advertising executives surveyed, 82% thought younger US consumers felt positive about AI-generated ads. Among the consumers, 45% reported feeling positive. That leaves a 37-percentage-point gap between the industry’s expectation and the audience’s answer.

The survey included 505 Gen Z and Millennial consumers and 104 advertising executives. Your audience may feel differently. Find out before treating an enthusiastic team meeting as approval from the people you want to reach.

Show the page to someone who fits your intended audience. Ask what they think they would receive, what they would pay and what could change that price. Let them find the answers. If you have to explain the offer aloud, write down what the page failed to explain.

The label and the work can get different reactions

AI content does not produce one predictable audience response. Consider two studies that looked at different parts of the problem.

A May 2026 paper in the Journal of Consumer Research examined AI disclosures on social media. Its observational TikTok models found roughly 7–8% fewer likes on AI-labelled posts at comparable view counts. The paper also included eight preregistered experiments examining disclosure and engagement. The authors identified perceived creator effort and connection with the creator as part of the explanation.

That matters when your pitch depends on expertise. A reader should be able to see what you contributed: the source you checked, the choice you made, the reason a recommendation fits their situation. Calling something expert advice leaves the reader looking for the expertise.

Meanwhile, the January 2025 working paper AI in disguise compared 4,633 display ads. After controls, it found no statistically significant average click-through difference between AI-generated and human-made ads. AI images perceived as less artificial could outperform human-made imagery.

So a blanket verdict on AI would miss the useful questions. Does the claim hold up? Does the work suit the audience? Can a reader see why they should trust it? Those questions deserve separate answers.

Read the promise before polishing the sentence

Start with the parts a customer could act on: prices, eligibility, delivery dates, refunds and promised results. Put the approved source beside each claim. If nobody can find one, pause that claim and ask whoever owns the offer.

Then follow the customer’s route. Open the advertisement, its destination page, the help page and the document available to download. Compare the conditions across them. Updating the landing page will not repair an older attachment already sitting in someone’s inbox.

Translation adds another place for meaning to change. Read the target language against the approved source, with a reviewer qualified in both languages and the subject. Check the amount, the unit, the deadline and the condition attached to the offer. Leave style preferences until those agree.

This is where context earns its place in the work. State who will read the text, what they need to understand and which names or terms must stay consistent. Reasoning Translator’s workflow lets you supply a translation brief for those instructions. Explore Reasoning Translator’s context-led translation workflow to see the public walkthrough.

Our translation brief guide provides a structure you can adapt. Keep the source, instructions and review decisions together. A reviewer should be able to trace a choice without having to reconstruct last week’s conversation.

Check what actually went live

Before publishing, save the version that was approved and name the person responsible for unresolved questions. If a price or condition changes afterward, have the affected passage checked again.

After publishing, open the page on a phone, follow its links and download the file. Compare what arrived with what was approved. Ask a reader to explain the offer in their own words. Record the misunderstanding and the change it prompted.

Catch contradictions before the campaign spreads. Once the promise has travelled through emails, ads and partner pages, finding every copy becomes another job. The customer should not have to discover which version you meant.

What the evidence leaves open

Sources checked 31 October 2026. The Air Canada award was damages, interest and fees, not a fine; the decision neither identifies the chatbot’s technology nor establishes that payment occurred. IAB measured US attitudes. The TikTok estimate is observational engagement evidence, and the advertising working paper concerns images on one platform without AI-origin disclosure. These results cannot supply a percentage of customers lost. A non-significant average also does not establish that every ad performs equally. We have not tested whether this review process increases sales or benchmarked Reasoning Translator here. The guide’s example is fictional preparation material. Qualified human review remains necessary, with legal, subject or cultural expertise appropriate to the document.